|
--- |
|
license: cc-by-4.0 |
|
datasets: |
|
- bltlab/queryner |
|
language: |
|
- en |
|
metrics: |
|
- f1 |
|
pipeline_tag: token-classification |
|
inference: |
|
parameters: |
|
aggregation_strategy: "first" |
|
--- |
|
|
|
# Model Card for Model ID |
|
|
|
E-commerce query segmentation model in English. |
|
This model is trained on QueryNER training dataset with the addition of augmentations so the model should be more robust to spelling mistakes and mentions unseen in the training data. |
|
|
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
This is a token classification model using BERT base uncased as the base model. |
|
The model is fine-tuned on the (QueryNER training dataset)[https://huggingface.co/datasets/bltlab/queryner] and augmented data as described in the QueryNER paper. |
|
|
|
|
|
- **Developed by:** [BLT Lab](https://github.com/bltlab) in collaboration with eBay. |
|
- **Funded by:** eBay |
|
- **Shared by:** (@cpalenmichel)[https://github.com/cpalenmichel] |
|
- **Model type:** Token Classification / Sequence Labeling / Chunking |
|
- **Language(s) (NLP):** English |
|
- **License:** CC-BY 4.0 |
|
- **Finetuned from model:** BERT base uncased |
|
|
|
### Model Sources |
|
|
|
Underlying model is based on [BERT base-uncased](https://huggingface.co/google-bert/bert-base-uncased). |
|
|
|
- **Repository:** [https://github.com/bltlab/query-ner](https://github.com/bltlab/query-ner) |
|
- **Paper:** Accepted at LREC-COLING Coming soon |
|
|
|
## Uses |
|
|
|
### Direct Use |
|
|
|
Intended use is research purposes and e-commerce query segmentation. |
|
|
|
### Downstream Use |
|
|
|
Potential downstream use cases include weighting entity spans, linking to knowledge bases, removing spans as a recovery strategy for null and low recall queries. |
|
|
|
### Out-of-Scope Use |
|
|
|
This model is trained only on the training data of the QueryNER dataset. It may not perform well on other domains without additional training data and further fine-tuning. |
|
|
|
## Bias, Risks, and Limitations |
|
|
|
See paper limitations section. |
|
|
|
## How to Get Started with the Model |
|
|
|
See huggingface tutorials for token classification and access the model using AutoModelForTokenClassification. |
|
Note that we do some post processing to make use of only the first subtoken's tag unlike the inference API. |
|
|
|
## Training Details |
|
|
|
### Training Data |
|
|
|
See paper for details. |
|
|
|
|
|
### Training Procedure |
|
|
|
See paper for details. |
|
|
|
#### Training Hyperparameters |
|
|
|
See paper for details. |
|
|
|
|
|
## Evaluation |
|
|
|
Evaluation details provided in the paper. |
|
Scoring was done using [SeqScore](https://github.com/bltlab/seqscore) using the conlleval repair method for invalid label transition sequences. |
|
|
|
### Testing Data, Factors & Metrics |
|
|
|
#### Testing Data |
|
|
|
QueryNER test set: [https://huggingface.co/datasets/bltlab/queryner](https://huggingface.co/datasets/bltlab/queryner) |
|
|
|
|
|
#### Factors |
|
Evaluation is reported with micro-F1 at the entity level on the QueryNER test set. |
|
We used conlleval repair method for invalid label transitions. |
|
|
|
#### Metrics |
|
We use micro-F1 at the entity level as this is fairly common practice for NER models. |
|
|
|
### Results |
|
|
|
[More Information Needed] |
|
|
|
|
|
## Environmental Impact |
|
Rough estimate |
|
|
|
- **Hardware Type:** 1 RTX 3090 GPU |
|
- **Hours used:** < 2 hours |
|
- **Cloud Provider:** Private |
|
- **Compute Region:** northamerica-northeast1 |
|
- **Carbon Emitted:** 0.02 |
|
|
|
|
|
## Citation |
|
|
|
Accepted at LREC-COLING coming soon |
|
|
|
**BibTeX:** |
|
|
|
Accepted at LREC-COLING coming soon |
|
|
|
|
|
## Model Card Authors |
|
|
|
Chester Palen-Michel (@cpalenmichel)[https://github.com/cpalenmichel] |
|
|
|
## Model Card Contact |
|
|
|
Chester Palen-Michel (@cpalenmichel)[https://github.com/cpalenmichel] |
